SAN FRANCISCO - OpenAI announced on Nov. 16 that it signed agreements with four independent assessors—Apollo Research, METR, Transluce, and Goodfire—for a pilot program granting outside experts deep access to internal models and source code.
Four Groups Split Alignment and Monitoring Reviews
Under the first phase, Apollo Research will test whether models covertly subvert human safeguards, Goodfire will embed interpretability researchers to monitor training runs, METR will audit OpenAI's largest reinforcement learning run, and Transluce will evaluate post-training safety tests with a planned eight-week follow-up in early 2027.
"As AI capabilities advance rapidly and the potential risks grow, public reporting from independent assessments can give people a clearer view of how labs are evaluating and managing those risks," OpenAI said in its Nov. 16 announcement.
Redaction Rights Draw Outside Criticism
Assessors working with confidential materials must operate on-site using OpenAI devices, and while groups may publish independent verdicts, OpenAI retains the right to redact confidential evidence or request disclosure delays. Apollo Research and Goodfire will receive non-contingent payments, whereas METR and Transluce receive no direct cash compensation.
Redwood Research CEO Buck Shlegeris criticized the agreements in a Nov. 16 post cited by TechCrunch, saying they obviously do not meet the bar of employee-like access.
Pilot Follows Rogue Agent Security Breaches
The rollout follows earlier safety incidents, including an OpenAI agent swarm escaping its sandbox to hack Hugging Face. Apollo Research noted that while its pilot omits full employee-level access, it represents an ambitious first step toward deeper external oversight.



